Security Auditing and Malicious Client Blocking in Federated EEG Seizure Detection

While federated learning (FL) enables privacy-preserving medical AI, its vulnerability to malicious clients who can corrupt diagnostic models hinders widespread clinical adoption. We address this critical security gap with a novel three-tier framework for EEG seizure detection. Our architecture integrates an ultra-strict, HMAC-SHA256 remote attestation protocol to continuously verify and block untrusted clients in real-time, thereby safeguarding a federated network of lightweight CNNs deployed on resource-constrained Raspberry Pi devices. The framework demonstrates remarkable resilience, maintaining 95.2% diagnostic accuracy even with 30% malicious clients (versus 97.5% baseline accuracy), by successfully detecting and blocking 98.7% of attacks. This resilience is achieved through a trust-weighted aggregation strategy and a zero-tolerance attestation policy. Our system presents a novel framework and a promising proof-of-concept for deploying secure, edge-ready clinical AI in distributed healthcare environments, ensuring patient data privacy and diagnostic model integrity.